Trustworthy AI

With rapidly advancing capabilities and adoption across commercial enterprise, sciences and public sector, there has never been a greater need for guardrails and controls.

Artificial intelligence (AI) has become an essential technology to respond to the complex challenges of today. Long an enabler of commerce and science behind the scenes, recent developments with foundation models, Generative AI and general-purpose application have cast AI into the limelight. Extended capabilities come with familiar risks, manifested in new ways. As AI permeates products, services, and infrastructure, it is critical to ensure its ethical application and quality implementation – following the principles of trustworthy AI such as fairness, transparency, accuracy, robustness, reliability, safety and security, as well as safeguarding privacy and confidentiality.

Artificial Intelligence (AI) has long fascinated both computer scientists, “futurists” and the public since the term was coined in the 1950s. Earlier expectations that AI could soon surpass human intelligence made for sensationalist headlines, but also for disappointment and the so-called “AI-Winter.” Since then, AI has emerged from the lab as a useful tool powering commerce and a wide variety of use cases across many other industries. Musings about runaway AIs gradually gave way to a grounded, realistic grasp of what AI is today: a set of sophisticated technologies with the potential to deliver significant economic, scientific and societal advantages. Experts estimate global economic impact of AI (including Generative AI) between $13 and $15,7 trillion by 2030, affecting productivity, innovation and increased consumer demand for AI-supported products and services.

Implemented properly, AI enables processes to run leaner & faster, products to be smarter, services more personalized – at scale. With AI, we can examine and learn from data at a pace and breadth that took our predecessors generations. Proper implementation is not automatic – it requires expertise, experience and effort. Open source toolkits have effectively “democratized” software development and led to a rapid proliferation in AI-based tools – from experts and debutants alike. This trend has only accelerated with the introduction of Generative AI, which adds creativity to a long list of AI capabilities. 

The Generative AI Explosion

The introduction of ChatGPT in 2023 was a major milestone for AI. It was not the first Large Language Model, not the first Generative AI (e.g., AI models that “create” new content). However, it was among the first to be so widely accessible and so easy to use, that it catapulted Generative AI – and indeed AI in general – into the mainstream. Another important feature of this milestone was the sudden dominance of “foundation models” – massively large deep neural networks (“transformers” and “diffusion models”) – so named for their unfathomable size and provision of base capabilities.  Where the development of foundation models is out of reach for most organizations, their developers have made them available – different versions to the public and to enterprise – encouraging users to apply them to all manner ways.

The “transformer” architecture at the heart of the Large Language Model was originally designed to predict the next word in a sentence, an ability which improved dramatically with ever larger training sets.  In addition to understanding natural language, the models also gained a command of general knowledge. With additional tuning, they could be crafted to follow instructions and answer questions, as we have become familiar with ChatGPT.  This ushered in a new concept – “general purpose AI.”  Where previous “narrow AI” models were limited to performing a single task (often exceedingly well), foundation models could perform admirably across wide range of tasks.  While “general purpose AI” does not equate to “general AI” (true intelligence), it is unarguably a major step in that direction. 

Yet with the growing list of AI capabilities come new permutations of familiar risks. Large Language Models have been known to “hallucinate” – that is, to provide factually incorrect, even nonsensical answers to questions with the utmost confidence and convincing. They can also inconsistently answer questions formulated in different ways; especially when taboo topics have been curated away, creative prompts can still eek out answers to unsavory questions. Generative AI models have a perennial struggle with copyrights, as their generated content can often closely correspond to training data which may be protected. Their ability to deep fake text, audio, image and video ventures into identity-theft territory, a flagrant violation of personal identity and privacy. And lastly, their multi-modal capabilities arm a vast number of bad actors with advanced tools to defraud others and execute cyber-attacks at a volume and level of sophistication against which few are prepared to defend.

Numerous design decisions affect the quality of an AI system

A common difference between AI (Machine Learning) and "classic" deterministic (rules-based) approaches is that AI learns from data rather than from a set of human-prescribed rules. This bestows a deceiving impression of objectivity. Supervised learning models – such as deep neural networks and the transformer architectures behind Generative AI – are only as good as the quality and scope of data on which they have been trained. Computer vision tasks conveniently illustrate the point. An algorithm is trained on a set of image data that are labeled for concepts, such as "stop sign" – by humans. The Deep Neural Network (DNN) classifies each image, breaks it down into characteristics (e.g., edges, colors, and shapes) and associates the result with the label. The DNN can do this very effectively. Yet even the best architectures fail if misled by training data, such as “stop sign” images given the “no entry” label. The resulting DNN will not be able to distinguish stop signs, either consistently or erratically assigning them to the “no entry” category. The AI is only as good as the human trained it, entirely dependent on the selection of data, its completeness, and consistently correct annotation. 

Similar to narrow AI, foundation models depend on annotation – labeling – to associate concepts with one another. Due to their sheer size, labeling has become a sophisticated and semi-automated science unto itself, with many clever innovations enabling their developers to imbue these enormous black box models with volumes of human knowledge. Ultimately, however, it is human beings who select the data, who perform the labeling, who invent methods to efficiently scale labeling. These amount to a plethora of design decisions, many of which can have profound impact on the functioning of an AI:

  • which data used to train? which labels?
  • with which objective function / question being asked?
  • utilizing which training approach?
  • which algorithmic architecture?
  • which tunings, tweaking and curation?

Implemented improperly, an AI model can systematically discriminate against what it does not know (what was absent from the training data), inadvertently perpetuating historical bias. It may succeed in classifying images for the wrong reasons (the background of the image vs the subject), a defect which cannot be caught without sufficient transparency. It may be unstable, making a prediction one way, then another, despite similar inputs. These risks are not new: they have accompanied model designers long before the “age of AI”. As statistical models, AI will nearly always give an answer, even if it is wrong or delivered with a low prediction confidence. 

To achieve the promise of AI, stakeholders must be ready to trust in their outputs

Practitioners accumulate best practices, optimized to suit whichever type of problem, nature of data, or desired solution functionality is on the agenda. While there is no one-size-fits-all checklist, these practices generally address risks and errors that have been encountered over the collective experience of developers, data scientists and machine learning engineers. Deloitte consolidated this experience and summarized into an easily understandable, yet wide-reaching “framework” of Trustworthy AI principles that capture the essence of what AI systems must fulfill to earn our trust.

Whether motivated by improving AI Quality or by compliance to upcoming regulations such as the EU AI Act or standards such as AI TRISM, these near universal Trustworthy AI principles provide useful orientation to AI practitioners and AI leaders alike. They apply as well to “classical” AI as they do to the newly popular “Generative AI” variant, but that’s another story…

Trustworthy AI Framework | How can it help you?

Find more relevant cases and information about trustworthy AI in you industry or sector.

Deep dives and relevant cases

open in new window Click here